What is training data?

Training data is the material, web pages, books, code, forum posts, and more, that gets fed to a model during the process that builds it. The model doesn't store this text the way a database stores rows; it adjusts its internal settings based on the patterns in it, then the raw text is gone. That's why a model can't reliably quote its training data verbatim or tell you exactly which pages it learned something from, what's left is a compressed sense of pattern, not a filing cabinet.

This distinction matters for how sites think about AI crawlers. A crawler like GPTBot collects pages for training data, so blocking it affects what a future version of a model might absorb, months or years later. It does nothing to whether that same page can be cited in an AI answer today, because live citations usually come from a separate search step at the moment of the question, not from training (see training-vs-search-crawler). Confusing the two leads to the common but wrong assumption that blocking a training crawler makes a site invisible to AI answers.